<s>
artificial	O
speech	O
from	O
text	O
(	O
text-to-speech	B-Application
)	O
or	O
spectrum	O
(	O
vocoder	O
)	O
.	O
</s>
<s>
a	O
text-to-speech	B-Application
system	O
,	O
the	O
associated	O
labels	O
and/or	O
input	O
text	O
.	O
</s>
<s>
Some	O
DNN-based	O
speech	B-Application
synthesizers	I-Application
are	O
approaching	O
the	O
naturalness	O
of	O
the	O
human	O
voice	O
.	O
</s>
<s>
The	O
acoustic	O
feature	O
is	O
typically	O
Spectrogram	O
or	O
spectrogram	O
in	O
Mel	B-Architecture
scale	I-Architecture
.	O
</s>
<s>
The	O
Mel-frequency	B-Algorithm
cepstrum	I-Algorithm
feature	O
used	O
in	O
the	O
speech	O
recognition	O
task	O
is	O
not	O
suitable	O
for	O
speech	B-Application
synthesis	I-Application
because	O
it	O
reduces	O
too	O
much	O
information	O
.	O
</s>
<s>
In	O
September	O
2016	O
,	O
DeepMind	B-Application
proposed	O
WaveNet	B-Application
,	O
a	O
deep	O
generative	O
model	O
of	O
raw	O
audio	O
waveforms	O
,	O
demonstrating	O
that	O
deep	O
learning-based	O
models	O
are	O
capable	O
of	O
modeling	O
raw	O
waveforms	O
and	O
generating	O
speech	O
from	O
acoustic	O
features	O
like	O
spectrograms	O
or	O
mel-spectrograms	B-Algorithm
.	O
</s>
<s>
Although	O
WaveNet	B-Application
was	O
initially	O
considered	O
to	O
be	O
computationally	O
expensive	O
and	O
slow	O
to	O
be	O
used	O
in	O
consumer	O
products	O
at	O
the	O
time	O
,	O
a	O
year	O
after	O
its	O
release	O
,	O
DeepMind	B-Application
unveiled	O
a	O
modified	O
version	O
of	O
WaveNet	B-Application
known	O
as	O
"	O
Parallel	O
WaveNet	B-Application
,	O
"	O
a	O
production	O
model	O
1,000	O
faster	O
than	O
the	O
original	O
.	O
</s>
<s>
In	O
the	O
same	O
year	O
,	O
Google	B-Application
and	O
Facebook	B-Application
proposed	O
and	O
,	O
respectively	O
,	O
to	O
generate	O
acoustic	O
features	O
directly	O
from	O
the	O
input	O
text	O
;	O
months	O
later	O
,	O
Google	B-Application
proposed	O
,	O
which	O
combined	O
the	O
WaveNet	B-Application
vocoder	O
with	O
the	O
revised	O
Tacotron	O
architecture	O
to	O
perform	O
end-to-end	O
speech	B-Application
synthesis	I-Application
.	O
</s>
<s>
Since	O
then	O
,	O
end-to-end	O
methods	O
have	O
become	O
the	O
hottest	O
research	O
topic	O
because	O
many	O
researchers	O
around	O
the	O
world	O
have	O
started	O
to	O
notice	O
the	O
power	O
of	O
end-to-end	O
speech	B-Application
synthesizers	I-Application
.	O
</s>
<s>
Currently	O
,	O
self-supervised	B-General_Concept
learning	I-General_Concept
has	O
gained	O
much	O
attention	O
through	O
better	O
use	O
of	O
unlabelled	O
data	O
.	O
</s>
<s>
In	O
June	O
2018	O
,	O
Google	B-Application
proposed	O
to	O
use	O
pre-trained	O
speaker	O
verification	O
models	O
as	O
speaker	O
encoders	O
to	O
extract	O
speaker	O
embeddings	O
.	O
</s>
<s>
The	O
speaker	O
encoders	O
then	O
become	O
part	O
of	O
the	O
neural	O
text-to-speech	B-Application
models	O
,	O
so	O
that	O
it	O
can	O
determine	O
the	O
style	O
and	O
characteristics	O
of	O
the	O
output	O
speech	O
.	O
</s>
<s>
In	O
deep	O
learning-based	O
speech	B-Application
synthesis	I-Application
,	O
neural	O
vocoders	O
play	O
an	O
important	O
role	O
in	O
generating	O
high-quality	O
speech	O
from	O
acoustic	O
features	O
.	O
</s>
<s>
The	O
WaveNet	B-Application
model	O
proposed	O
in	O
2016	O
achieves	O
excellent	O
performance	O
on	O
speech	O
quality	O
.	O
</s>
<s>
However	O
,	O
the	O
auto-regressive	O
nature	O
of	O
WaveNet	B-Application
makes	O
the	O
inference	O
process	O
dramatically	O
slow	O
.	O
</s>
<s>
To	O
solve	O
this	O
problem	O
,	O
Parallel	O
WaveNet	B-Application
was	O
proposed	O
.	O
</s>
<s>
Parallel	O
WaveNet	B-Application
is	O
an	O
inverse	O
autoregressive	O
flow-based	O
model	O
which	O
is	O
trained	O
by	O
knowledge	B-Algorithm
distillation	I-Algorithm
with	O
a	O
pre-trained	O
teacher	O
WaveNet	B-Application
model	O
.	O
</s>
<s>
However	O
,	O
despite	O
the	O
high	O
inference	O
speed	O
,	O
parallel	O
WaveNet	B-Application
has	O
the	O
limitation	O
of	O
needing	O
a	O
pre-trained	O
WaveNet	B-Application
model	O
,	O
so	O
that	O
WaveGlow	O
takes	O
many	O
weeks	O
to	O
converge	O
with	O
limited	O
computing	O
devices	O
.	O
</s>
